btrfs: Remove V0 extent support
[sfrench/cifs-2.6.git] / fs / btrfs / ctree.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
4  */
5
6 #include <linux/sched.h>
7 #include <linux/slab.h>
8 #include <linux/rbtree.h>
9 #include <linux/mm.h>
10 #include "ctree.h"
11 #include "disk-io.h"
12 #include "transaction.h"
13 #include "print-tree.h"
14 #include "locking.h"
15
16 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
17                       *root, struct btrfs_path *path, int level);
18 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
19                       const struct btrfs_key *ins_key, struct btrfs_path *path,
20                       int data_size, int extend);
21 static int push_node_left(struct btrfs_trans_handle *trans,
22                           struct btrfs_fs_info *fs_info,
23                           struct extent_buffer *dst,
24                           struct extent_buffer *src, int empty);
25 static int balance_node_right(struct btrfs_trans_handle *trans,
26                               struct btrfs_fs_info *fs_info,
27                               struct extent_buffer *dst_buf,
28                               struct extent_buffer *src_buf);
29 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
30                     int level, int slot);
31
32 struct btrfs_path *btrfs_alloc_path(void)
33 {
34         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
35 }
36
37 /*
38  * set all locked nodes in the path to blocking locks.  This should
39  * be done before scheduling
40  */
41 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
42 {
43         int i;
44         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
45                 if (!p->nodes[i] || !p->locks[i])
46                         continue;
47                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
48                 if (p->locks[i] == BTRFS_READ_LOCK)
49                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
50                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
51                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
52         }
53 }
54
55 /*
56  * reset all the locked nodes in the patch to spinning locks.
57  *
58  * held is used to keep lockdep happy, when lockdep is enabled
59  * we set held to a blocking lock before we go around and
60  * retake all the spinlocks in the path.  You can safely use NULL
61  * for held
62  */
63 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
64                                         struct extent_buffer *held, int held_rw)
65 {
66         int i;
67
68         if (held) {
69                 btrfs_set_lock_blocking_rw(held, held_rw);
70                 if (held_rw == BTRFS_WRITE_LOCK)
71                         held_rw = BTRFS_WRITE_LOCK_BLOCKING;
72                 else if (held_rw == BTRFS_READ_LOCK)
73                         held_rw = BTRFS_READ_LOCK_BLOCKING;
74         }
75         btrfs_set_path_blocking(p);
76
77         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
78                 if (p->nodes[i] && p->locks[i]) {
79                         btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
80                         if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
81                                 p->locks[i] = BTRFS_WRITE_LOCK;
82                         else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
83                                 p->locks[i] = BTRFS_READ_LOCK;
84                 }
85         }
86
87         if (held)
88                 btrfs_clear_lock_blocking_rw(held, held_rw);
89 }
90
91 /* this also releases the path */
92 void btrfs_free_path(struct btrfs_path *p)
93 {
94         if (!p)
95                 return;
96         btrfs_release_path(p);
97         kmem_cache_free(btrfs_path_cachep, p);
98 }
99
100 /*
101  * path release drops references on the extent buffers in the path
102  * and it drops any locks held by this path
103  *
104  * It is safe to call this on paths that no locks or extent buffers held.
105  */
106 noinline void btrfs_release_path(struct btrfs_path *p)
107 {
108         int i;
109
110         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
111                 p->slots[i] = 0;
112                 if (!p->nodes[i])
113                         continue;
114                 if (p->locks[i]) {
115                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
116                         p->locks[i] = 0;
117                 }
118                 free_extent_buffer(p->nodes[i]);
119                 p->nodes[i] = NULL;
120         }
121 }
122
123 /*
124  * safely gets a reference on the root node of a tree.  A lock
125  * is not taken, so a concurrent writer may put a different node
126  * at the root of the tree.  See btrfs_lock_root_node for the
127  * looping required.
128  *
129  * The extent buffer returned by this has a reference taken, so
130  * it won't disappear.  It may stop being the root of the tree
131  * at any time because there are no locks held.
132  */
133 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
134 {
135         struct extent_buffer *eb;
136
137         while (1) {
138                 rcu_read_lock();
139                 eb = rcu_dereference(root->node);
140
141                 /*
142                  * RCU really hurts here, we could free up the root node because
143                  * it was COWed but we may not get the new root node yet so do
144                  * the inc_not_zero dance and if it doesn't work then
145                  * synchronize_rcu and try again.
146                  */
147                 if (atomic_inc_not_zero(&eb->refs)) {
148                         rcu_read_unlock();
149                         break;
150                 }
151                 rcu_read_unlock();
152                 synchronize_rcu();
153         }
154         return eb;
155 }
156
157 /* loop around taking references on and locking the root node of the
158  * tree until you end up with a lock on the root.  A locked buffer
159  * is returned, with a reference held.
160  */
161 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
162 {
163         struct extent_buffer *eb;
164
165         while (1) {
166                 eb = btrfs_root_node(root);
167                 btrfs_tree_lock(eb);
168                 if (eb == root->node)
169                         break;
170                 btrfs_tree_unlock(eb);
171                 free_extent_buffer(eb);
172         }
173         return eb;
174 }
175
176 /* loop around taking references on and locking the root node of the
177  * tree until you end up with a lock on the root.  A locked buffer
178  * is returned, with a reference held.
179  */
180 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
181 {
182         struct extent_buffer *eb;
183
184         while (1) {
185                 eb = btrfs_root_node(root);
186                 btrfs_tree_read_lock(eb);
187                 if (eb == root->node)
188                         break;
189                 btrfs_tree_read_unlock(eb);
190                 free_extent_buffer(eb);
191         }
192         return eb;
193 }
194
195 /* cowonly root (everything not a reference counted cow subvolume), just get
196  * put onto a simple dirty list.  transaction.c walks this to make sure they
197  * get properly updated on disk.
198  */
199 static void add_root_to_dirty_list(struct btrfs_root *root)
200 {
201         struct btrfs_fs_info *fs_info = root->fs_info;
202
203         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
204             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
205                 return;
206
207         spin_lock(&fs_info->trans_lock);
208         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
209                 /* Want the extent tree to be the last on the list */
210                 if (root->objectid == BTRFS_EXTENT_TREE_OBJECTID)
211                         list_move_tail(&root->dirty_list,
212                                        &fs_info->dirty_cowonly_roots);
213                 else
214                         list_move(&root->dirty_list,
215                                   &fs_info->dirty_cowonly_roots);
216         }
217         spin_unlock(&fs_info->trans_lock);
218 }
219
220 /*
221  * used by snapshot creation to make a copy of a root for a tree with
222  * a given objectid.  The buffer with the new root node is returned in
223  * cow_ret, and this func returns zero on success or a negative error code.
224  */
225 int btrfs_copy_root(struct btrfs_trans_handle *trans,
226                       struct btrfs_root *root,
227                       struct extent_buffer *buf,
228                       struct extent_buffer **cow_ret, u64 new_root_objectid)
229 {
230         struct btrfs_fs_info *fs_info = root->fs_info;
231         struct extent_buffer *cow;
232         int ret = 0;
233         int level;
234         struct btrfs_disk_key disk_key;
235
236         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
237                 trans->transid != fs_info->running_transaction->transid);
238         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
239                 trans->transid != root->last_trans);
240
241         level = btrfs_header_level(buf);
242         if (level == 0)
243                 btrfs_item_key(buf, &disk_key, 0);
244         else
245                 btrfs_node_key(buf, &disk_key, 0);
246
247         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
248                         &disk_key, level, buf->start, 0);
249         if (IS_ERR(cow))
250                 return PTR_ERR(cow);
251
252         copy_extent_buffer_full(cow, buf);
253         btrfs_set_header_bytenr(cow, cow->start);
254         btrfs_set_header_generation(cow, trans->transid);
255         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
256         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
257                                      BTRFS_HEADER_FLAG_RELOC);
258         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
259                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
260         else
261                 btrfs_set_header_owner(cow, new_root_objectid);
262
263         write_extent_buffer_fsid(cow, fs_info->fsid);
264
265         WARN_ON(btrfs_header_generation(buf) > trans->transid);
266         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
267                 ret = btrfs_inc_ref(trans, root, cow, 1);
268         else
269                 ret = btrfs_inc_ref(trans, root, cow, 0);
270
271         if (ret)
272                 return ret;
273
274         btrfs_mark_buffer_dirty(cow);
275         *cow_ret = cow;
276         return 0;
277 }
278
279 enum mod_log_op {
280         MOD_LOG_KEY_REPLACE,
281         MOD_LOG_KEY_ADD,
282         MOD_LOG_KEY_REMOVE,
283         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
284         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
285         MOD_LOG_MOVE_KEYS,
286         MOD_LOG_ROOT_REPLACE,
287 };
288
289 struct tree_mod_root {
290         u64 logical;
291         u8 level;
292 };
293
294 struct tree_mod_elem {
295         struct rb_node node;
296         u64 logical;
297         u64 seq;
298         enum mod_log_op op;
299
300         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
301         int slot;
302
303         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
304         u64 generation;
305
306         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
307         struct btrfs_disk_key key;
308         u64 blockptr;
309
310         /* this is used for op == MOD_LOG_MOVE_KEYS */
311         struct {
312                 int dst_slot;
313                 int nr_items;
314         } move;
315
316         /* this is used for op == MOD_LOG_ROOT_REPLACE */
317         struct tree_mod_root old_root;
318 };
319
320 /*
321  * Pull a new tree mod seq number for our operation.
322  */
323 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
324 {
325         return atomic64_inc_return(&fs_info->tree_mod_seq);
326 }
327
328 /*
329  * This adds a new blocker to the tree mod log's blocker list if the @elem
330  * passed does not already have a sequence number set. So when a caller expects
331  * to record tree modifications, it should ensure to set elem->seq to zero
332  * before calling btrfs_get_tree_mod_seq.
333  * Returns a fresh, unused tree log modification sequence number, even if no new
334  * blocker was added.
335  */
336 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
337                            struct seq_list *elem)
338 {
339         write_lock(&fs_info->tree_mod_log_lock);
340         spin_lock(&fs_info->tree_mod_seq_lock);
341         if (!elem->seq) {
342                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
343                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
344         }
345         spin_unlock(&fs_info->tree_mod_seq_lock);
346         write_unlock(&fs_info->tree_mod_log_lock);
347
348         return elem->seq;
349 }
350
351 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
352                             struct seq_list *elem)
353 {
354         struct rb_root *tm_root;
355         struct rb_node *node;
356         struct rb_node *next;
357         struct seq_list *cur_elem;
358         struct tree_mod_elem *tm;
359         u64 min_seq = (u64)-1;
360         u64 seq_putting = elem->seq;
361
362         if (!seq_putting)
363                 return;
364
365         spin_lock(&fs_info->tree_mod_seq_lock);
366         list_del(&elem->list);
367         elem->seq = 0;
368
369         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
370                 if (cur_elem->seq < min_seq) {
371                         if (seq_putting > cur_elem->seq) {
372                                 /*
373                                  * blocker with lower sequence number exists, we
374                                  * cannot remove anything from the log
375                                  */
376                                 spin_unlock(&fs_info->tree_mod_seq_lock);
377                                 return;
378                         }
379                         min_seq = cur_elem->seq;
380                 }
381         }
382         spin_unlock(&fs_info->tree_mod_seq_lock);
383
384         /*
385          * anything that's lower than the lowest existing (read: blocked)
386          * sequence number can be removed from the tree.
387          */
388         write_lock(&fs_info->tree_mod_log_lock);
389         tm_root = &fs_info->tree_mod_log;
390         for (node = rb_first(tm_root); node; node = next) {
391                 next = rb_next(node);
392                 tm = rb_entry(node, struct tree_mod_elem, node);
393                 if (tm->seq > min_seq)
394                         continue;
395                 rb_erase(node, tm_root);
396                 kfree(tm);
397         }
398         write_unlock(&fs_info->tree_mod_log_lock);
399 }
400
401 /*
402  * key order of the log:
403  *       node/leaf start address -> sequence
404  *
405  * The 'start address' is the logical address of the *new* root node
406  * for root replace operations, or the logical address of the affected
407  * block for all other operations.
408  *
409  * Note: must be called with write lock for fs_info::tree_mod_log_lock.
410  */
411 static noinline int
412 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
413 {
414         struct rb_root *tm_root;
415         struct rb_node **new;
416         struct rb_node *parent = NULL;
417         struct tree_mod_elem *cur;
418
419         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
420
421         tm_root = &fs_info->tree_mod_log;
422         new = &tm_root->rb_node;
423         while (*new) {
424                 cur = rb_entry(*new, struct tree_mod_elem, node);
425                 parent = *new;
426                 if (cur->logical < tm->logical)
427                         new = &((*new)->rb_left);
428                 else if (cur->logical > tm->logical)
429                         new = &((*new)->rb_right);
430                 else if (cur->seq < tm->seq)
431                         new = &((*new)->rb_left);
432                 else if (cur->seq > tm->seq)
433                         new = &((*new)->rb_right);
434                 else
435                         return -EEXIST;
436         }
437
438         rb_link_node(&tm->node, parent, new);
439         rb_insert_color(&tm->node, tm_root);
440         return 0;
441 }
442
443 /*
444  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
445  * returns zero with the tree_mod_log_lock acquired. The caller must hold
446  * this until all tree mod log insertions are recorded in the rb tree and then
447  * write unlock fs_info::tree_mod_log_lock.
448  */
449 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
450                                     struct extent_buffer *eb) {
451         smp_mb();
452         if (list_empty(&(fs_info)->tree_mod_seq_list))
453                 return 1;
454         if (eb && btrfs_header_level(eb) == 0)
455                 return 1;
456
457         write_lock(&fs_info->tree_mod_log_lock);
458         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
459                 write_unlock(&fs_info->tree_mod_log_lock);
460                 return 1;
461         }
462
463         return 0;
464 }
465
466 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
467 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
468                                     struct extent_buffer *eb)
469 {
470         smp_mb();
471         if (list_empty(&(fs_info)->tree_mod_seq_list))
472                 return 0;
473         if (eb && btrfs_header_level(eb) == 0)
474                 return 0;
475
476         return 1;
477 }
478
479 static struct tree_mod_elem *
480 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
481                     enum mod_log_op op, gfp_t flags)
482 {
483         struct tree_mod_elem *tm;
484
485         tm = kzalloc(sizeof(*tm), flags);
486         if (!tm)
487                 return NULL;
488
489         tm->logical = eb->start;
490         if (op != MOD_LOG_KEY_ADD) {
491                 btrfs_node_key(eb, &tm->key, slot);
492                 tm->blockptr = btrfs_node_blockptr(eb, slot);
493         }
494         tm->op = op;
495         tm->slot = slot;
496         tm->generation = btrfs_node_ptr_generation(eb, slot);
497         RB_CLEAR_NODE(&tm->node);
498
499         return tm;
500 }
501
502 static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
503                 enum mod_log_op op, gfp_t flags)
504 {
505         struct tree_mod_elem *tm;
506         int ret;
507
508         if (!tree_mod_need_log(eb->fs_info, eb))
509                 return 0;
510
511         tm = alloc_tree_mod_elem(eb, slot, op, flags);
512         if (!tm)
513                 return -ENOMEM;
514
515         if (tree_mod_dont_log(eb->fs_info, eb)) {
516                 kfree(tm);
517                 return 0;
518         }
519
520         ret = __tree_mod_log_insert(eb->fs_info, tm);
521         write_unlock(&eb->fs_info->tree_mod_log_lock);
522         if (ret)
523                 kfree(tm);
524
525         return ret;
526 }
527
528 static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
529                 int dst_slot, int src_slot, int nr_items)
530 {
531         struct tree_mod_elem *tm = NULL;
532         struct tree_mod_elem **tm_list = NULL;
533         int ret = 0;
534         int i;
535         int locked = 0;
536
537         if (!tree_mod_need_log(eb->fs_info, eb))
538                 return 0;
539
540         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
541         if (!tm_list)
542                 return -ENOMEM;
543
544         tm = kzalloc(sizeof(*tm), GFP_NOFS);
545         if (!tm) {
546                 ret = -ENOMEM;
547                 goto free_tms;
548         }
549
550         tm->logical = eb->start;
551         tm->slot = src_slot;
552         tm->move.dst_slot = dst_slot;
553         tm->move.nr_items = nr_items;
554         tm->op = MOD_LOG_MOVE_KEYS;
555
556         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
557                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
558                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
559                 if (!tm_list[i]) {
560                         ret = -ENOMEM;
561                         goto free_tms;
562                 }
563         }
564
565         if (tree_mod_dont_log(eb->fs_info, eb))
566                 goto free_tms;
567         locked = 1;
568
569         /*
570          * When we override something during the move, we log these removals.
571          * This can only happen when we move towards the beginning of the
572          * buffer, i.e. dst_slot < src_slot.
573          */
574         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
575                 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
576                 if (ret)
577                         goto free_tms;
578         }
579
580         ret = __tree_mod_log_insert(eb->fs_info, tm);
581         if (ret)
582                 goto free_tms;
583         write_unlock(&eb->fs_info->tree_mod_log_lock);
584         kfree(tm_list);
585
586         return 0;
587 free_tms:
588         for (i = 0; i < nr_items; i++) {
589                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
590                         rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
591                 kfree(tm_list[i]);
592         }
593         if (locked)
594                 write_unlock(&eb->fs_info->tree_mod_log_lock);
595         kfree(tm_list);
596         kfree(tm);
597
598         return ret;
599 }
600
601 static inline int
602 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
603                        struct tree_mod_elem **tm_list,
604                        int nritems)
605 {
606         int i, j;
607         int ret;
608
609         for (i = nritems - 1; i >= 0; i--) {
610                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
611                 if (ret) {
612                         for (j = nritems - 1; j > i; j--)
613                                 rb_erase(&tm_list[j]->node,
614                                          &fs_info->tree_mod_log);
615                         return ret;
616                 }
617         }
618
619         return 0;
620 }
621
622 static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
623                          struct extent_buffer *new_root, int log_removal)
624 {
625         struct btrfs_fs_info *fs_info = old_root->fs_info;
626         struct tree_mod_elem *tm = NULL;
627         struct tree_mod_elem **tm_list = NULL;
628         int nritems = 0;
629         int ret = 0;
630         int i;
631
632         if (!tree_mod_need_log(fs_info, NULL))
633                 return 0;
634
635         if (log_removal && btrfs_header_level(old_root) > 0) {
636                 nritems = btrfs_header_nritems(old_root);
637                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
638                                   GFP_NOFS);
639                 if (!tm_list) {
640                         ret = -ENOMEM;
641                         goto free_tms;
642                 }
643                 for (i = 0; i < nritems; i++) {
644                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
645                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
646                         if (!tm_list[i]) {
647                                 ret = -ENOMEM;
648                                 goto free_tms;
649                         }
650                 }
651         }
652
653         tm = kzalloc(sizeof(*tm), GFP_NOFS);
654         if (!tm) {
655                 ret = -ENOMEM;
656                 goto free_tms;
657         }
658
659         tm->logical = new_root->start;
660         tm->old_root.logical = old_root->start;
661         tm->old_root.level = btrfs_header_level(old_root);
662         tm->generation = btrfs_header_generation(old_root);
663         tm->op = MOD_LOG_ROOT_REPLACE;
664
665         if (tree_mod_dont_log(fs_info, NULL))
666                 goto free_tms;
667
668         if (tm_list)
669                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
670         if (!ret)
671                 ret = __tree_mod_log_insert(fs_info, tm);
672
673         write_unlock(&fs_info->tree_mod_log_lock);
674         if (ret)
675                 goto free_tms;
676         kfree(tm_list);
677
678         return ret;
679
680 free_tms:
681         if (tm_list) {
682                 for (i = 0; i < nritems; i++)
683                         kfree(tm_list[i]);
684                 kfree(tm_list);
685         }
686         kfree(tm);
687
688         return ret;
689 }
690
691 static struct tree_mod_elem *
692 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
693                       int smallest)
694 {
695         struct rb_root *tm_root;
696         struct rb_node *node;
697         struct tree_mod_elem *cur = NULL;
698         struct tree_mod_elem *found = NULL;
699
700         read_lock(&fs_info->tree_mod_log_lock);
701         tm_root = &fs_info->tree_mod_log;
702         node = tm_root->rb_node;
703         while (node) {
704                 cur = rb_entry(node, struct tree_mod_elem, node);
705                 if (cur->logical < start) {
706                         node = node->rb_left;
707                 } else if (cur->logical > start) {
708                         node = node->rb_right;
709                 } else if (cur->seq < min_seq) {
710                         node = node->rb_left;
711                 } else if (!smallest) {
712                         /* we want the node with the highest seq */
713                         if (found)
714                                 BUG_ON(found->seq > cur->seq);
715                         found = cur;
716                         node = node->rb_left;
717                 } else if (cur->seq > min_seq) {
718                         /* we want the node with the smallest seq */
719                         if (found)
720                                 BUG_ON(found->seq < cur->seq);
721                         found = cur;
722                         node = node->rb_right;
723                 } else {
724                         found = cur;
725                         break;
726                 }
727         }
728         read_unlock(&fs_info->tree_mod_log_lock);
729
730         return found;
731 }
732
733 /*
734  * this returns the element from the log with the smallest time sequence
735  * value that's in the log (the oldest log item). any element with a time
736  * sequence lower than min_seq will be ignored.
737  */
738 static struct tree_mod_elem *
739 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
740                            u64 min_seq)
741 {
742         return __tree_mod_log_search(fs_info, start, min_seq, 1);
743 }
744
745 /*
746  * this returns the element from the log with the largest time sequence
747  * value that's in the log (the most recent log item). any element with
748  * a time sequence lower than min_seq will be ignored.
749  */
750 static struct tree_mod_elem *
751 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
752 {
753         return __tree_mod_log_search(fs_info, start, min_seq, 0);
754 }
755
756 static noinline int
757 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
758                      struct extent_buffer *src, unsigned long dst_offset,
759                      unsigned long src_offset, int nr_items)
760 {
761         int ret = 0;
762         struct tree_mod_elem **tm_list = NULL;
763         struct tree_mod_elem **tm_list_add, **tm_list_rem;
764         int i;
765         int locked = 0;
766
767         if (!tree_mod_need_log(fs_info, NULL))
768                 return 0;
769
770         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
771                 return 0;
772
773         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
774                           GFP_NOFS);
775         if (!tm_list)
776                 return -ENOMEM;
777
778         tm_list_add = tm_list;
779         tm_list_rem = tm_list + nr_items;
780         for (i = 0; i < nr_items; i++) {
781                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
782                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
783                 if (!tm_list_rem[i]) {
784                         ret = -ENOMEM;
785                         goto free_tms;
786                 }
787
788                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
789                     MOD_LOG_KEY_ADD, GFP_NOFS);
790                 if (!tm_list_add[i]) {
791                         ret = -ENOMEM;
792                         goto free_tms;
793                 }
794         }
795
796         if (tree_mod_dont_log(fs_info, NULL))
797                 goto free_tms;
798         locked = 1;
799
800         for (i = 0; i < nr_items; i++) {
801                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
802                 if (ret)
803                         goto free_tms;
804                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
805                 if (ret)
806                         goto free_tms;
807         }
808
809         write_unlock(&fs_info->tree_mod_log_lock);
810         kfree(tm_list);
811
812         return 0;
813
814 free_tms:
815         for (i = 0; i < nr_items * 2; i++) {
816                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
817                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
818                 kfree(tm_list[i]);
819         }
820         if (locked)
821                 write_unlock(&fs_info->tree_mod_log_lock);
822         kfree(tm_list);
823
824         return ret;
825 }
826
827 static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
828 {
829         struct tree_mod_elem **tm_list = NULL;
830         int nritems = 0;
831         int i;
832         int ret = 0;
833
834         if (btrfs_header_level(eb) == 0)
835                 return 0;
836
837         if (!tree_mod_need_log(eb->fs_info, NULL))
838                 return 0;
839
840         nritems = btrfs_header_nritems(eb);
841         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
842         if (!tm_list)
843                 return -ENOMEM;
844
845         for (i = 0; i < nritems; i++) {
846                 tm_list[i] = alloc_tree_mod_elem(eb, i,
847                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
848                 if (!tm_list[i]) {
849                         ret = -ENOMEM;
850                         goto free_tms;
851                 }
852         }
853
854         if (tree_mod_dont_log(eb->fs_info, eb))
855                 goto free_tms;
856
857         ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
858         write_unlock(&eb->fs_info->tree_mod_log_lock);
859         if (ret)
860                 goto free_tms;
861         kfree(tm_list);
862
863         return 0;
864
865 free_tms:
866         for (i = 0; i < nritems; i++)
867                 kfree(tm_list[i]);
868         kfree(tm_list);
869
870         return ret;
871 }
872
873 /*
874  * check if the tree block can be shared by multiple trees
875  */
876 int btrfs_block_can_be_shared(struct btrfs_root *root,
877                               struct extent_buffer *buf)
878 {
879         /*
880          * Tree blocks not in reference counted trees and tree roots
881          * are never shared. If a block was allocated after the last
882          * snapshot and the block was not allocated by tree relocation,
883          * we know the block is not shared.
884          */
885         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
886             buf != root->node && buf != root->commit_root &&
887             (btrfs_header_generation(buf) <=
888              btrfs_root_last_snapshot(&root->root_item) ||
889              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
890                 return 1;
891
892         return 0;
893 }
894
895 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
896                                        struct btrfs_root *root,
897                                        struct extent_buffer *buf,
898                                        struct extent_buffer *cow,
899                                        int *last_ref)
900 {
901         struct btrfs_fs_info *fs_info = root->fs_info;
902         u64 refs;
903         u64 owner;
904         u64 flags;
905         u64 new_flags = 0;
906         int ret;
907
908         /*
909          * Backrefs update rules:
910          *
911          * Always use full backrefs for extent pointers in tree block
912          * allocated by tree relocation.
913          *
914          * If a shared tree block is no longer referenced by its owner
915          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
916          * use full backrefs for extent pointers in tree block.
917          *
918          * If a tree block is been relocating
919          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
920          * use full backrefs for extent pointers in tree block.
921          * The reason for this is some operations (such as drop tree)
922          * are only allowed for blocks use full backrefs.
923          */
924
925         if (btrfs_block_can_be_shared(root, buf)) {
926                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
927                                                btrfs_header_level(buf), 1,
928                                                &refs, &flags);
929                 if (ret)
930                         return ret;
931                 if (refs == 0) {
932                         ret = -EROFS;
933                         btrfs_handle_fs_error(fs_info, ret, NULL);
934                         return ret;
935                 }
936         } else {
937                 refs = 1;
938                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
939                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
940                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
941                 else
942                         flags = 0;
943         }
944
945         owner = btrfs_header_owner(buf);
946         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
947                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
948
949         if (refs > 1) {
950                 if ((owner == root->root_key.objectid ||
951                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
952                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
953                         ret = btrfs_inc_ref(trans, root, buf, 1);
954                         if (ret)
955                                 return ret;
956
957                         if (root->root_key.objectid ==
958                             BTRFS_TREE_RELOC_OBJECTID) {
959                                 ret = btrfs_dec_ref(trans, root, buf, 0);
960                                 if (ret)
961                                         return ret;
962                                 ret = btrfs_inc_ref(trans, root, cow, 1);
963                                 if (ret)
964                                         return ret;
965                         }
966                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
967                 } else {
968
969                         if (root->root_key.objectid ==
970                             BTRFS_TREE_RELOC_OBJECTID)
971                                 ret = btrfs_inc_ref(trans, root, cow, 1);
972                         else
973                                 ret = btrfs_inc_ref(trans, root, cow, 0);
974                         if (ret)
975                                 return ret;
976                 }
977                 if (new_flags != 0) {
978                         int level = btrfs_header_level(buf);
979
980                         ret = btrfs_set_disk_extent_flags(trans, fs_info,
981                                                           buf->start,
982                                                           buf->len,
983                                                           new_flags, level, 0);
984                         if (ret)
985                                 return ret;
986                 }
987         } else {
988                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
989                         if (root->root_key.objectid ==
990                             BTRFS_TREE_RELOC_OBJECTID)
991                                 ret = btrfs_inc_ref(trans, root, cow, 1);
992                         else
993                                 ret = btrfs_inc_ref(trans, root, cow, 0);
994                         if (ret)
995                                 return ret;
996                         ret = btrfs_dec_ref(trans, root, buf, 1);
997                         if (ret)
998                                 return ret;
999                 }
1000                 clean_tree_block(fs_info, buf);
1001                 *last_ref = 1;
1002         }
1003         return 0;
1004 }
1005
1006 /*
1007  * does the dirty work in cow of a single block.  The parent block (if
1008  * supplied) is updated to point to the new cow copy.  The new buffer is marked
1009  * dirty and returned locked.  If you modify the block it needs to be marked
1010  * dirty again.
1011  *
1012  * search_start -- an allocation hint for the new block
1013  *
1014  * empty_size -- a hint that you plan on doing more cow.  This is the size in
1015  * bytes the allocator should try to find free next to the block it returns.
1016  * This is just a hint and may be ignored by the allocator.
1017  */
1018 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1019                              struct btrfs_root *root,
1020                              struct extent_buffer *buf,
1021                              struct extent_buffer *parent, int parent_slot,
1022                              struct extent_buffer **cow_ret,
1023                              u64 search_start, u64 empty_size)
1024 {
1025         struct btrfs_fs_info *fs_info = root->fs_info;
1026         struct btrfs_disk_key disk_key;
1027         struct extent_buffer *cow;
1028         int level, ret;
1029         int last_ref = 0;
1030         int unlock_orig = 0;
1031         u64 parent_start = 0;
1032
1033         if (*cow_ret == buf)
1034                 unlock_orig = 1;
1035
1036         btrfs_assert_tree_locked(buf);
1037
1038         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1039                 trans->transid != fs_info->running_transaction->transid);
1040         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1041                 trans->transid != root->last_trans);
1042
1043         level = btrfs_header_level(buf);
1044
1045         if (level == 0)
1046                 btrfs_item_key(buf, &disk_key, 0);
1047         else
1048                 btrfs_node_key(buf, &disk_key, 0);
1049
1050         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1051                 parent_start = parent->start;
1052
1053         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1054                         root->root_key.objectid, &disk_key, level,
1055                         search_start, empty_size);
1056         if (IS_ERR(cow))
1057                 return PTR_ERR(cow);
1058
1059         /* cow is set to blocking by btrfs_init_new_buffer */
1060
1061         copy_extent_buffer_full(cow, buf);
1062         btrfs_set_header_bytenr(cow, cow->start);
1063         btrfs_set_header_generation(cow, trans->transid);
1064         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1065         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1066                                      BTRFS_HEADER_FLAG_RELOC);
1067         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1068                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1069         else
1070                 btrfs_set_header_owner(cow, root->root_key.objectid);
1071
1072         write_extent_buffer_fsid(cow, fs_info->fsid);
1073
1074         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1075         if (ret) {
1076                 btrfs_abort_transaction(trans, ret);
1077                 return ret;
1078         }
1079
1080         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1081                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1082                 if (ret) {
1083                         btrfs_abort_transaction(trans, ret);
1084                         return ret;
1085                 }
1086         }
1087
1088         if (buf == root->node) {
1089                 WARN_ON(parent && parent != buf);
1090                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1091                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1092                         parent_start = buf->start;
1093
1094                 extent_buffer_get(cow);
1095                 ret = tree_mod_log_insert_root(root->node, cow, 1);
1096                 BUG_ON(ret < 0);
1097                 rcu_assign_pointer(root->node, cow);
1098
1099                 btrfs_free_tree_block(trans, root, buf, parent_start,
1100                                       last_ref);
1101                 free_extent_buffer(buf);
1102                 add_root_to_dirty_list(root);
1103         } else {
1104                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1105                 tree_mod_log_insert_key(parent, parent_slot,
1106                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1107                 btrfs_set_node_blockptr(parent, parent_slot,
1108                                         cow->start);
1109                 btrfs_set_node_ptr_generation(parent, parent_slot,
1110                                               trans->transid);
1111                 btrfs_mark_buffer_dirty(parent);
1112                 if (last_ref) {
1113                         ret = tree_mod_log_free_eb(buf);
1114                         if (ret) {
1115                                 btrfs_abort_transaction(trans, ret);
1116                                 return ret;
1117                         }
1118                 }
1119                 btrfs_free_tree_block(trans, root, buf, parent_start,
1120                                       last_ref);
1121         }
1122         if (unlock_orig)
1123                 btrfs_tree_unlock(buf);
1124         free_extent_buffer_stale(buf);
1125         btrfs_mark_buffer_dirty(cow);
1126         *cow_ret = cow;
1127         return 0;
1128 }
1129
1130 /*
1131  * returns the logical address of the oldest predecessor of the given root.
1132  * entries older than time_seq are ignored.
1133  */
1134 static struct tree_mod_elem *__tree_mod_log_oldest_root(
1135                 struct extent_buffer *eb_root, u64 time_seq)
1136 {
1137         struct tree_mod_elem *tm;
1138         struct tree_mod_elem *found = NULL;
1139         u64 root_logical = eb_root->start;
1140         int looped = 0;
1141
1142         if (!time_seq)
1143                 return NULL;
1144
1145         /*
1146          * the very last operation that's logged for a root is the
1147          * replacement operation (if it is replaced at all). this has
1148          * the logical address of the *new* root, making it the very
1149          * first operation that's logged for this root.
1150          */
1151         while (1) {
1152                 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1153                                                 time_seq);
1154                 if (!looped && !tm)
1155                         return NULL;
1156                 /*
1157                  * if there are no tree operation for the oldest root, we simply
1158                  * return it. this should only happen if that (old) root is at
1159                  * level 0.
1160                  */
1161                 if (!tm)
1162                         break;
1163
1164                 /*
1165                  * if there's an operation that's not a root replacement, we
1166                  * found the oldest version of our root. normally, we'll find a
1167                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1168                  */
1169                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1170                         break;
1171
1172                 found = tm;
1173                 root_logical = tm->old_root.logical;
1174                 looped = 1;
1175         }
1176
1177         /* if there's no old root to return, return what we found instead */
1178         if (!found)
1179                 found = tm;
1180
1181         return found;
1182 }
1183
1184 /*
1185  * tm is a pointer to the first operation to rewind within eb. then, all
1186  * previous operations will be rewound (until we reach something older than
1187  * time_seq).
1188  */
1189 static void
1190 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1191                       u64 time_seq, struct tree_mod_elem *first_tm)
1192 {
1193         u32 n;
1194         struct rb_node *next;
1195         struct tree_mod_elem *tm = first_tm;
1196         unsigned long o_dst;
1197         unsigned long o_src;
1198         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1199
1200         n = btrfs_header_nritems(eb);
1201         read_lock(&fs_info->tree_mod_log_lock);
1202         while (tm && tm->seq >= time_seq) {
1203                 /*
1204                  * all the operations are recorded with the operator used for
1205                  * the modification. as we're going backwards, we do the
1206                  * opposite of each operation here.
1207                  */
1208                 switch (tm->op) {
1209                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1210                         BUG_ON(tm->slot < n);
1211                         /* Fallthrough */
1212                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1213                 case MOD_LOG_KEY_REMOVE:
1214                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1215                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1216                         btrfs_set_node_ptr_generation(eb, tm->slot,
1217                                                       tm->generation);
1218                         n++;
1219                         break;
1220                 case MOD_LOG_KEY_REPLACE:
1221                         BUG_ON(tm->slot >= n);
1222                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1223                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1224                         btrfs_set_node_ptr_generation(eb, tm->slot,
1225                                                       tm->generation);
1226                         break;
1227                 case MOD_LOG_KEY_ADD:
1228                         /* if a move operation is needed it's in the log */
1229                         n--;
1230                         break;
1231                 case MOD_LOG_MOVE_KEYS:
1232                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1233                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1234                         memmove_extent_buffer(eb, o_dst, o_src,
1235                                               tm->move.nr_items * p_size);
1236                         break;
1237                 case MOD_LOG_ROOT_REPLACE:
1238                         /*
1239                          * this operation is special. for roots, this must be
1240                          * handled explicitly before rewinding.
1241                          * for non-roots, this operation may exist if the node
1242                          * was a root: root A -> child B; then A gets empty and
1243                          * B is promoted to the new root. in the mod log, we'll
1244                          * have a root-replace operation for B, a tree block
1245                          * that is no root. we simply ignore that operation.
1246                          */
1247                         break;
1248                 }
1249                 next = rb_next(&tm->node);
1250                 if (!next)
1251                         break;
1252                 tm = rb_entry(next, struct tree_mod_elem, node);
1253                 if (tm->logical != first_tm->logical)
1254                         break;
1255         }
1256         read_unlock(&fs_info->tree_mod_log_lock);
1257         btrfs_set_header_nritems(eb, n);
1258 }
1259
1260 /*
1261  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1262  * is returned. If rewind operations happen, a fresh buffer is returned. The
1263  * returned buffer is always read-locked. If the returned buffer is not the
1264  * input buffer, the lock on the input buffer is released and the input buffer
1265  * is freed (its refcount is decremented).
1266  */
1267 static struct extent_buffer *
1268 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1269                     struct extent_buffer *eb, u64 time_seq)
1270 {
1271         struct extent_buffer *eb_rewin;
1272         struct tree_mod_elem *tm;
1273
1274         if (!time_seq)
1275                 return eb;
1276
1277         if (btrfs_header_level(eb) == 0)
1278                 return eb;
1279
1280         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1281         if (!tm)
1282                 return eb;
1283
1284         btrfs_set_path_blocking(path);
1285         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1286
1287         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1288                 BUG_ON(tm->slot != 0);
1289                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1290                 if (!eb_rewin) {
1291                         btrfs_tree_read_unlock_blocking(eb);
1292                         free_extent_buffer(eb);
1293                         return NULL;
1294                 }
1295                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1296                 btrfs_set_header_backref_rev(eb_rewin,
1297                                              btrfs_header_backref_rev(eb));
1298                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1299                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1300         } else {
1301                 eb_rewin = btrfs_clone_extent_buffer(eb);
1302                 if (!eb_rewin) {
1303                         btrfs_tree_read_unlock_blocking(eb);
1304                         free_extent_buffer(eb);
1305                         return NULL;
1306                 }
1307         }
1308
1309         btrfs_clear_path_blocking(path, NULL, BTRFS_READ_LOCK);
1310         btrfs_tree_read_unlock_blocking(eb);
1311         free_extent_buffer(eb);
1312
1313         extent_buffer_get(eb_rewin);
1314         btrfs_tree_read_lock(eb_rewin);
1315         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1316         WARN_ON(btrfs_header_nritems(eb_rewin) >
1317                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1318
1319         return eb_rewin;
1320 }
1321
1322 /*
1323  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1324  * value. If there are no changes, the current root->root_node is returned. If
1325  * anything changed in between, there's a fresh buffer allocated on which the
1326  * rewind operations are done. In any case, the returned buffer is read locked.
1327  * Returns NULL on error (with no locks held).
1328  */
1329 static inline struct extent_buffer *
1330 get_old_root(struct btrfs_root *root, u64 time_seq)
1331 {
1332         struct btrfs_fs_info *fs_info = root->fs_info;
1333         struct tree_mod_elem *tm;
1334         struct extent_buffer *eb = NULL;
1335         struct extent_buffer *eb_root;
1336         struct extent_buffer *old;
1337         struct tree_mod_root *old_root = NULL;
1338         u64 old_generation = 0;
1339         u64 logical;
1340         int level;
1341
1342         eb_root = btrfs_read_lock_root_node(root);
1343         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1344         if (!tm)
1345                 return eb_root;
1346
1347         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1348                 old_root = &tm->old_root;
1349                 old_generation = tm->generation;
1350                 logical = old_root->logical;
1351                 level = old_root->level;
1352         } else {
1353                 logical = eb_root->start;
1354                 level = btrfs_header_level(eb_root);
1355         }
1356
1357         tm = tree_mod_log_search(fs_info, logical, time_seq);
1358         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1359                 btrfs_tree_read_unlock(eb_root);
1360                 free_extent_buffer(eb_root);
1361                 old = read_tree_block(fs_info, logical, 0, level, NULL);
1362                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1363                         if (!IS_ERR(old))
1364                                 free_extent_buffer(old);
1365                         btrfs_warn(fs_info,
1366                                    "failed to read tree block %llu from get_old_root",
1367                                    logical);
1368                 } else {
1369                         eb = btrfs_clone_extent_buffer(old);
1370                         free_extent_buffer(old);
1371                 }
1372         } else if (old_root) {
1373                 btrfs_tree_read_unlock(eb_root);
1374                 free_extent_buffer(eb_root);
1375                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1376         } else {
1377                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1378                 eb = btrfs_clone_extent_buffer(eb_root);
1379                 btrfs_tree_read_unlock_blocking(eb_root);
1380                 free_extent_buffer(eb_root);
1381         }
1382
1383         if (!eb)
1384                 return NULL;
1385         extent_buffer_get(eb);
1386         btrfs_tree_read_lock(eb);
1387         if (old_root) {
1388                 btrfs_set_header_bytenr(eb, eb->start);
1389                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1390                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1391                 btrfs_set_header_level(eb, old_root->level);
1392                 btrfs_set_header_generation(eb, old_generation);
1393         }
1394         if (tm)
1395                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1396         else
1397                 WARN_ON(btrfs_header_level(eb) != 0);
1398         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1399
1400         return eb;
1401 }
1402
1403 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1404 {
1405         struct tree_mod_elem *tm;
1406         int level;
1407         struct extent_buffer *eb_root = btrfs_root_node(root);
1408
1409         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1410         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1411                 level = tm->old_root.level;
1412         } else {
1413                 level = btrfs_header_level(eb_root);
1414         }
1415         free_extent_buffer(eb_root);
1416
1417         return level;
1418 }
1419
1420 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1421                                    struct btrfs_root *root,
1422                                    struct extent_buffer *buf)
1423 {
1424         if (btrfs_is_testing(root->fs_info))
1425                 return 0;
1426
1427         /* Ensure we can see the FORCE_COW bit */
1428         smp_mb__before_atomic();
1429
1430         /*
1431          * We do not need to cow a block if
1432          * 1) this block is not created or changed in this transaction;
1433          * 2) this block does not belong to TREE_RELOC tree;
1434          * 3) the root is not forced COW.
1435          *
1436          * What is forced COW:
1437          *    when we create snapshot during committing the transaction,
1438          *    after we've finished coping src root, we must COW the shared
1439          *    block to ensure the metadata consistency.
1440          */
1441         if (btrfs_header_generation(buf) == trans->transid &&
1442             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1443             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1444               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1445             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1446                 return 0;
1447         return 1;
1448 }
1449
1450 /*
1451  * cows a single block, see __btrfs_cow_block for the real work.
1452  * This version of it has extra checks so that a block isn't COWed more than
1453  * once per transaction, as long as it hasn't been written yet
1454  */
1455 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1456                     struct btrfs_root *root, struct extent_buffer *buf,
1457                     struct extent_buffer *parent, int parent_slot,
1458                     struct extent_buffer **cow_ret)
1459 {
1460         struct btrfs_fs_info *fs_info = root->fs_info;
1461         u64 search_start;
1462         int ret;
1463
1464         if (trans->transaction != fs_info->running_transaction)
1465                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1466                        trans->transid,
1467                        fs_info->running_transaction->transid);
1468
1469         if (trans->transid != fs_info->generation)
1470                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1471                        trans->transid, fs_info->generation);
1472
1473         if (!should_cow_block(trans, root, buf)) {
1474                 trans->dirty = true;
1475                 *cow_ret = buf;
1476                 return 0;
1477         }
1478
1479         search_start = buf->start & ~((u64)SZ_1G - 1);
1480
1481         if (parent)
1482                 btrfs_set_lock_blocking(parent);
1483         btrfs_set_lock_blocking(buf);
1484
1485         ret = __btrfs_cow_block(trans, root, buf, parent,
1486                                  parent_slot, cow_ret, search_start, 0);
1487
1488         trace_btrfs_cow_block(root, buf, *cow_ret);
1489
1490         return ret;
1491 }
1492
1493 /*
1494  * helper function for defrag to decide if two blocks pointed to by a
1495  * node are actually close by
1496  */
1497 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1498 {
1499         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1500                 return 1;
1501         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1502                 return 1;
1503         return 0;
1504 }
1505
1506 /*
1507  * compare two keys in a memcmp fashion
1508  */
1509 static int comp_keys(const struct btrfs_disk_key *disk,
1510                      const struct btrfs_key *k2)
1511 {
1512         struct btrfs_key k1;
1513
1514         btrfs_disk_key_to_cpu(&k1, disk);
1515
1516         return btrfs_comp_cpu_keys(&k1, k2);
1517 }
1518
1519 /*
1520  * same as comp_keys only with two btrfs_key's
1521  */
1522 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1523 {
1524         if (k1->objectid > k2->objectid)
1525                 return 1;
1526         if (k1->objectid < k2->objectid)
1527                 return -1;
1528         if (k1->type > k2->type)
1529                 return 1;
1530         if (k1->type < k2->type)
1531                 return -1;
1532         if (k1->offset > k2->offset)
1533                 return 1;
1534         if (k1->offset < k2->offset)
1535                 return -1;
1536         return 0;
1537 }
1538
1539 /*
1540  * this is used by the defrag code to go through all the
1541  * leaves pointed to by a node and reallocate them so that
1542  * disk order is close to key order
1543  */
1544 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1545                        struct btrfs_root *root, struct extent_buffer *parent,
1546                        int start_slot, u64 *last_ret,
1547                        struct btrfs_key *progress)
1548 {
1549         struct btrfs_fs_info *fs_info = root->fs_info;
1550         struct extent_buffer *cur;
1551         u64 blocknr;
1552         u64 gen;
1553         u64 search_start = *last_ret;
1554         u64 last_block = 0;
1555         u64 other;
1556         u32 parent_nritems;
1557         int end_slot;
1558         int i;
1559         int err = 0;
1560         int parent_level;
1561         int uptodate;
1562         u32 blocksize;
1563         int progress_passed = 0;
1564         struct btrfs_disk_key disk_key;
1565
1566         parent_level = btrfs_header_level(parent);
1567
1568         WARN_ON(trans->transaction != fs_info->running_transaction);
1569         WARN_ON(trans->transid != fs_info->generation);
1570
1571         parent_nritems = btrfs_header_nritems(parent);
1572         blocksize = fs_info->nodesize;
1573         end_slot = parent_nritems - 1;
1574
1575         if (parent_nritems <= 1)
1576                 return 0;
1577
1578         btrfs_set_lock_blocking(parent);
1579
1580         for (i = start_slot; i <= end_slot; i++) {
1581                 struct btrfs_key first_key;
1582                 int close = 1;
1583
1584                 btrfs_node_key(parent, &disk_key, i);
1585                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1586                         continue;
1587
1588                 progress_passed = 1;
1589                 blocknr = btrfs_node_blockptr(parent, i);
1590                 gen = btrfs_node_ptr_generation(parent, i);
1591                 btrfs_node_key_to_cpu(parent, &first_key, i);
1592                 if (last_block == 0)
1593                         last_block = blocknr;
1594
1595                 if (i > 0) {
1596                         other = btrfs_node_blockptr(parent, i - 1);
1597                         close = close_blocks(blocknr, other, blocksize);
1598                 }
1599                 if (!close && i < end_slot) {
1600                         other = btrfs_node_blockptr(parent, i + 1);
1601                         close = close_blocks(blocknr, other, blocksize);
1602                 }
1603                 if (close) {
1604                         last_block = blocknr;
1605                         continue;
1606                 }
1607
1608                 cur = find_extent_buffer(fs_info, blocknr);
1609                 if (cur)
1610                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1611                 else
1612                         uptodate = 0;
1613                 if (!cur || !uptodate) {
1614                         if (!cur) {
1615                                 cur = read_tree_block(fs_info, blocknr, gen,
1616                                                       parent_level - 1,
1617                                                       &first_key);
1618                                 if (IS_ERR(cur)) {
1619                                         return PTR_ERR(cur);
1620                                 } else if (!extent_buffer_uptodate(cur)) {
1621                                         free_extent_buffer(cur);
1622                                         return -EIO;
1623                                 }
1624                         } else if (!uptodate) {
1625                                 err = btrfs_read_buffer(cur, gen,
1626                                                 parent_level - 1,&first_key);
1627                                 if (err) {
1628                                         free_extent_buffer(cur);
1629                                         return err;
1630                                 }
1631                         }
1632                 }
1633                 if (search_start == 0)
1634                         search_start = last_block;
1635
1636                 btrfs_tree_lock(cur);
1637                 btrfs_set_lock_blocking(cur);
1638                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1639                                         &cur, search_start,
1640                                         min(16 * blocksize,
1641                                             (end_slot - i) * blocksize));
1642                 if (err) {
1643                         btrfs_tree_unlock(cur);
1644                         free_extent_buffer(cur);
1645                         break;
1646                 }
1647                 search_start = cur->start;
1648                 last_block = cur->start;
1649                 *last_ret = search_start;
1650                 btrfs_tree_unlock(cur);
1651                 free_extent_buffer(cur);
1652         }
1653         return err;
1654 }
1655
1656 /*
1657  * search for key in the extent_buffer.  The items start at offset p,
1658  * and they are item_size apart.  There are 'max' items in p.
1659  *
1660  * the slot in the array is returned via slot, and it points to
1661  * the place where you would insert key if it is not found in
1662  * the array.
1663  *
1664  * slot may point to max if the key is bigger than all of the keys
1665  */
1666 static noinline int generic_bin_search(struct extent_buffer *eb,
1667                                        unsigned long p, int item_size,
1668                                        const struct btrfs_key *key,
1669                                        int max, int *slot)
1670 {
1671         int low = 0;
1672         int high = max;
1673         int mid;
1674         int ret;
1675         struct btrfs_disk_key *tmp = NULL;
1676         struct btrfs_disk_key unaligned;
1677         unsigned long offset;
1678         char *kaddr = NULL;
1679         unsigned long map_start = 0;
1680         unsigned long map_len = 0;
1681         int err;
1682
1683         if (low > high) {
1684                 btrfs_err(eb->fs_info,
1685                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1686                           __func__, low, high, eb->start,
1687                           btrfs_header_owner(eb), btrfs_header_level(eb));
1688                 return -EINVAL;
1689         }
1690
1691         while (low < high) {
1692                 mid = (low + high) / 2;
1693                 offset = p + mid * item_size;
1694
1695                 if (!kaddr || offset < map_start ||
1696                     (offset + sizeof(struct btrfs_disk_key)) >
1697                     map_start + map_len) {
1698
1699                         err = map_private_extent_buffer(eb, offset,
1700                                                 sizeof(struct btrfs_disk_key),
1701                                                 &kaddr, &map_start, &map_len);
1702
1703                         if (!err) {
1704                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1705                                                         map_start);
1706                         } else if (err == 1) {
1707                                 read_extent_buffer(eb, &unaligned,
1708                                                    offset, sizeof(unaligned));
1709                                 tmp = &unaligned;
1710                         } else {
1711                                 return err;
1712                         }
1713
1714                 } else {
1715                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1716                                                         map_start);
1717                 }
1718                 ret = comp_keys(tmp, key);
1719
1720                 if (ret < 0)
1721                         low = mid + 1;
1722                 else if (ret > 0)
1723                         high = mid;
1724                 else {
1725                         *slot = mid;
1726                         return 0;
1727                 }
1728         }
1729         *slot = low;
1730         return 1;
1731 }
1732
1733 /*
1734  * simple bin_search frontend that does the right thing for
1735  * leaves vs nodes
1736  */
1737 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1738                      int level, int *slot)
1739 {
1740         if (level == 0)
1741                 return generic_bin_search(eb,
1742                                           offsetof(struct btrfs_leaf, items),
1743                                           sizeof(struct btrfs_item),
1744                                           key, btrfs_header_nritems(eb),
1745                                           slot);
1746         else
1747                 return generic_bin_search(eb,
1748                                           offsetof(struct btrfs_node, ptrs),
1749                                           sizeof(struct btrfs_key_ptr),
1750                                           key, btrfs_header_nritems(eb),
1751                                           slot);
1752 }
1753
1754 static void root_add_used(struct btrfs_root *root, u32 size)
1755 {
1756         spin_lock(&root->accounting_lock);
1757         btrfs_set_root_used(&root->root_item,
1758                             btrfs_root_used(&root->root_item) + size);
1759         spin_unlock(&root->accounting_lock);
1760 }
1761
1762 static void root_sub_used(struct btrfs_root *root, u32 size)
1763 {
1764         spin_lock(&root->accounting_lock);
1765         btrfs_set_root_used(&root->root_item,
1766                             btrfs_root_used(&root->root_item) - size);
1767         spin_unlock(&root->accounting_lock);
1768 }
1769
1770 /* given a node and slot number, this reads the blocks it points to.  The
1771  * extent buffer is returned with a reference taken (but unlocked).
1772  */
1773 static noinline struct extent_buffer *
1774 read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1775                int slot)
1776 {
1777         int level = btrfs_header_level(parent);
1778         struct extent_buffer *eb;
1779         struct btrfs_key first_key;
1780
1781         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1782                 return ERR_PTR(-ENOENT);
1783
1784         BUG_ON(level == 0);
1785
1786         btrfs_node_key_to_cpu(parent, &first_key, slot);
1787         eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1788                              btrfs_node_ptr_generation(parent, slot),
1789                              level - 1, &first_key);
1790         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1791                 free_extent_buffer(eb);
1792                 eb = ERR_PTR(-EIO);
1793         }
1794
1795         return eb;
1796 }
1797
1798 /*
1799  * node level balancing, used to make sure nodes are in proper order for
1800  * item deletion.  We balance from the top down, so we have to make sure
1801  * that a deletion won't leave an node completely empty later on.
1802  */
1803 static noinline int balance_level(struct btrfs_trans_handle *trans,
1804                          struct btrfs_root *root,
1805                          struct btrfs_path *path, int level)
1806 {
1807         struct btrfs_fs_info *fs_info = root->fs_info;
1808         struct extent_buffer *right = NULL;
1809         struct extent_buffer *mid;
1810         struct extent_buffer *left = NULL;
1811         struct extent_buffer *parent = NULL;
1812         int ret = 0;
1813         int wret;
1814         int pslot;
1815         int orig_slot = path->slots[level];
1816         u64 orig_ptr;
1817
1818         if (level == 0)
1819                 return 0;
1820
1821         mid = path->nodes[level];
1822
1823         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1824                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1825         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1826
1827         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1828
1829         if (level < BTRFS_MAX_LEVEL - 1) {
1830                 parent = path->nodes[level + 1];
1831                 pslot = path->slots[level + 1];
1832         }
1833
1834         /*
1835          * deal with the case where there is only one pointer in the root
1836          * by promoting the node below to a root
1837          */
1838         if (!parent) {
1839                 struct extent_buffer *child;
1840
1841                 if (btrfs_header_nritems(mid) != 1)
1842                         return 0;
1843
1844                 /* promote the child to a root */
1845                 child = read_node_slot(fs_info, mid, 0);
1846                 if (IS_ERR(child)) {
1847                         ret = PTR_ERR(child);
1848                         btrfs_handle_fs_error(fs_info, ret, NULL);
1849                         goto enospc;
1850                 }
1851
1852                 btrfs_tree_lock(child);
1853                 btrfs_set_lock_blocking(child);
1854                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1855                 if (ret) {
1856                         btrfs_tree_unlock(child);
1857                         free_extent_buffer(child);
1858                         goto enospc;
1859                 }
1860
1861                 ret = tree_mod_log_insert_root(root->node, child, 1);
1862                 BUG_ON(ret < 0);
1863                 rcu_assign_pointer(root->node, child);
1864
1865                 add_root_to_dirty_list(root);
1866                 btrfs_tree_unlock(child);
1867
1868                 path->locks[level] = 0;
1869                 path->nodes[level] = NULL;
1870                 clean_tree_block(fs_info, mid);
1871                 btrfs_tree_unlock(mid);
1872                 /* once for the path */
1873                 free_extent_buffer(mid);
1874
1875                 root_sub_used(root, mid->len);
1876                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1877                 /* once for the root ptr */
1878                 free_extent_buffer_stale(mid);
1879                 return 0;
1880         }
1881         if (btrfs_header_nritems(mid) >
1882             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1883                 return 0;
1884
1885         left = read_node_slot(fs_info, parent, pslot - 1);
1886         if (IS_ERR(left))
1887                 left = NULL;
1888
1889         if (left) {
1890                 btrfs_tree_lock(left);
1891                 btrfs_set_lock_blocking(left);
1892                 wret = btrfs_cow_block(trans, root, left,
1893                                        parent, pslot - 1, &left);
1894                 if (wret) {
1895                         ret = wret;
1896                         goto enospc;
1897                 }
1898         }
1899
1900         right = read_node_slot(fs_info, parent, pslot + 1);
1901         if (IS_ERR(right))
1902                 right = NULL;
1903
1904         if (right) {
1905                 btrfs_tree_lock(right);
1906                 btrfs_set_lock_blocking(right);
1907                 wret = btrfs_cow_block(trans, root, right,
1908                                        parent, pslot + 1, &right);
1909                 if (wret) {
1910                         ret = wret;
1911                         goto enospc;
1912                 }
1913         }
1914
1915         /* first, try to make some room in the middle buffer */
1916         if (left) {
1917                 orig_slot += btrfs_header_nritems(left);
1918                 wret = push_node_left(trans, fs_info, left, mid, 1);
1919                 if (wret < 0)
1920                         ret = wret;
1921         }
1922
1923         /*
1924          * then try to empty the right most buffer into the middle
1925          */
1926         if (right) {
1927                 wret = push_node_left(trans, fs_info, mid, right, 1);
1928                 if (wret < 0 && wret != -ENOSPC)
1929                         ret = wret;
1930                 if (btrfs_header_nritems(right) == 0) {
1931                         clean_tree_block(fs_info, right);
1932                         btrfs_tree_unlock(right);
1933                         del_ptr(root, path, level + 1, pslot + 1);
1934                         root_sub_used(root, right->len);
1935                         btrfs_free_tree_block(trans, root, right, 0, 1);
1936                         free_extent_buffer_stale(right);
1937                         right = NULL;
1938                 } else {
1939                         struct btrfs_disk_key right_key;
1940                         btrfs_node_key(right, &right_key, 0);
1941                         ret = tree_mod_log_insert_key(parent, pslot + 1,
1942                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1943                         BUG_ON(ret < 0);
1944                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1945                         btrfs_mark_buffer_dirty(parent);
1946                 }
1947         }
1948         if (btrfs_header_nritems(mid) == 1) {
1949                 /*
1950                  * we're not allowed to leave a node with one item in the
1951                  * tree during a delete.  A deletion from lower in the tree
1952                  * could try to delete the only pointer in this node.
1953                  * So, pull some keys from the left.
1954                  * There has to be a left pointer at this point because
1955                  * otherwise we would have pulled some pointers from the
1956                  * right
1957                  */
1958                 if (!left) {
1959                         ret = -EROFS;
1960                         btrfs_handle_fs_error(fs_info, ret, NULL);
1961                         goto enospc;
1962                 }
1963                 wret = balance_node_right(trans, fs_info, mid, left);
1964                 if (wret < 0) {
1965                         ret = wret;
1966                         goto enospc;
1967                 }
1968                 if (wret == 1) {
1969                         wret = push_node_left(trans, fs_info, left, mid, 1);
1970                         if (wret < 0)
1971                                 ret = wret;
1972                 }
1973                 BUG_ON(wret == 1);
1974         }
1975         if (btrfs_header_nritems(mid) == 0) {
1976                 clean_tree_block(fs_info, mid);
1977                 btrfs_tree_unlock(mid);
1978                 del_ptr(root, path, level + 1, pslot);
1979                 root_sub_used(root, mid->len);
1980                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1981                 free_extent_buffer_stale(mid);
1982                 mid = NULL;
1983         } else {
1984                 /* update the parent key to reflect our changes */
1985                 struct btrfs_disk_key mid_key;
1986                 btrfs_node_key(mid, &mid_key, 0);
1987                 ret = tree_mod_log_insert_key(parent, pslot,
1988                                 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1989                 BUG_ON(ret < 0);
1990                 btrfs_set_node_key(parent, &mid_key, pslot);
1991                 btrfs_mark_buffer_dirty(parent);
1992         }
1993
1994         /* update the path */
1995         if (left) {
1996                 if (btrfs_header_nritems(left) > orig_slot) {
1997                         extent_buffer_get(left);
1998                         /* left was locked after cow */
1999                         path->nodes[level] = left;
2000                         path->slots[level + 1] -= 1;
2001                         path->slots[level] = orig_slot;
2002                         if (mid) {
2003                                 btrfs_tree_unlock(mid);
2004                                 free_extent_buffer(mid);
2005                         }
2006                 } else {
2007                         orig_slot -= btrfs_header_nritems(left);
2008                         path->slots[level] = orig_slot;
2009                 }
2010         }
2011         /* double check we haven't messed things up */
2012         if (orig_ptr !=
2013             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2014                 BUG();
2015 enospc:
2016         if (right) {
2017                 btrfs_tree_unlock(right);
2018                 free_extent_buffer(right);
2019         }
2020         if (left) {
2021                 if (path->nodes[level] != left)
2022                         btrfs_tree_unlock(left);
2023                 free_extent_buffer(left);
2024         }
2025         return ret;
2026 }
2027
2028 /* Node balancing for insertion.  Here we only split or push nodes around
2029  * when they are completely full.  This is also done top down, so we
2030  * have to be pessimistic.
2031  */
2032 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2033                                           struct btrfs_root *root,
2034                                           struct btrfs_path *path, int level)
2035 {
2036         struct btrfs_fs_info *fs_info = root->fs_info;
2037         struct extent_buffer *right = NULL;
2038         struct extent_buffer *mid;
2039         struct extent_buffer *left = NULL;
2040         struct extent_buffer *parent = NULL;
2041         int ret = 0;
2042         int wret;
2043         int pslot;
2044         int orig_slot = path->slots[level];
2045
2046         if (level == 0)
2047                 return 1;
2048
2049         mid = path->nodes[level];
2050         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2051
2052         if (level < BTRFS_MAX_LEVEL - 1) {
2053                 parent = path->nodes[level + 1];
2054                 pslot = path->slots[level + 1];
2055         }
2056
2057         if (!parent)
2058                 return 1;
2059
2060         left = read_node_slot(fs_info, parent, pslot - 1);
2061         if (IS_ERR(left))
2062                 left = NULL;
2063
2064         /* first, try to make some room in the middle buffer */
2065         if (left) {
2066                 u32 left_nr;
2067
2068                 btrfs_tree_lock(left);
2069                 btrfs_set_lock_blocking(left);
2070
2071                 left_nr = btrfs_header_nritems(left);
2072                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2073                         wret = 1;
2074                 } else {
2075                         ret = btrfs_cow_block(trans, root, left, parent,
2076                                               pslot - 1, &left);
2077                         if (ret)
2078                                 wret = 1;
2079                         else {
2080                                 wret = push_node_left(trans, fs_info,
2081                                                       left, mid, 0);
2082                         }
2083                 }
2084                 if (wret < 0)
2085                         ret = wret;
2086                 if (wret == 0) {
2087                         struct btrfs_disk_key disk_key;
2088                         orig_slot += left_nr;
2089                         btrfs_node_key(mid, &disk_key, 0);
2090                         ret = tree_mod_log_insert_key(parent, pslot,
2091                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2092                         BUG_ON(ret < 0);
2093                         btrfs_set_node_key(parent, &disk_key, pslot);
2094                         btrfs_mark_buffer_dirty(parent);
2095                         if (btrfs_header_nritems(left) > orig_slot) {
2096                                 path->nodes[level] = left;
2097                                 path->slots[level + 1] -= 1;
2098                                 path->slots[level] = orig_slot;
2099                                 btrfs_tree_unlock(mid);
2100                                 free_extent_buffer(mid);
2101                         } else {
2102                                 orig_slot -=
2103                                         btrfs_header_nritems(left);
2104                                 path->slots[level] = orig_slot;
2105                                 btrfs_tree_unlock(left);
2106                                 free_extent_buffer(left);
2107                         }
2108                         return 0;
2109                 }
2110                 btrfs_tree_unlock(left);
2111                 free_extent_buffer(left);
2112         }
2113         right = read_node_slot(fs_info, parent, pslot + 1);
2114         if (IS_ERR(right))
2115                 right = NULL;
2116
2117         /*
2118          * then try to empty the right most buffer into the middle
2119          */
2120         if (right) {
2121                 u32 right_nr;
2122
2123                 btrfs_tree_lock(right);
2124                 btrfs_set_lock_blocking(right);
2125
2126                 right_nr = btrfs_header_nritems(right);
2127                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2128                         wret = 1;
2129                 } else {
2130                         ret = btrfs_cow_block(trans, root, right,
2131                                               parent, pslot + 1,
2132                                               &right);
2133                         if (ret)
2134                                 wret = 1;
2135                         else {
2136                                 wret = balance_node_right(trans, fs_info,
2137                                                           right, mid);
2138                         }
2139                 }
2140                 if (wret < 0)
2141                         ret = wret;
2142                 if (wret == 0) {
2143                         struct btrfs_disk_key disk_key;
2144
2145                         btrfs_node_key(right, &disk_key, 0);
2146                         ret = tree_mod_log_insert_key(parent, pslot + 1,
2147                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2148                         BUG_ON(ret < 0);
2149                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2150                         btrfs_mark_buffer_dirty(parent);
2151
2152                         if (btrfs_header_nritems(mid) <= orig_slot) {
2153                                 path->nodes[level] = right;
2154                                 path->slots[level + 1] += 1;
2155                                 path->slots[level] = orig_slot -
2156                                         btrfs_header_nritems(mid);
2157                                 btrfs_tree_unlock(mid);
2158                                 free_extent_buffer(mid);
2159                         } else {
2160                                 btrfs_tree_unlock(right);
2161                                 free_extent_buffer(right);
2162                         }
2163                         return 0;
2164                 }
2165                 btrfs_tree_unlock(right);
2166                 free_extent_buffer(right);
2167         }
2168         return 1;
2169 }
2170
2171 /*
2172  * readahead one full node of leaves, finding things that are close
2173  * to the block in 'slot', and triggering ra on them.
2174  */
2175 static void reada_for_search(struct btrfs_fs_info *fs_info,
2176                              struct btrfs_path *path,
2177                              int level, int slot, u64 objectid)
2178 {
2179         struct extent_buffer *node;
2180         struct btrfs_disk_key disk_key;
2181         u32 nritems;
2182         u64 search;
2183         u64 target;
2184         u64 nread = 0;
2185         struct extent_buffer *eb;
2186         u32 nr;
2187         u32 blocksize;
2188         u32 nscan = 0;
2189
2190         if (level != 1)
2191                 return;
2192
2193         if (!path->nodes[level])
2194                 return;
2195
2196         node = path->nodes[level];
2197
2198         search = btrfs_node_blockptr(node, slot);
2199         blocksize = fs_info->nodesize;
2200         eb = find_extent_buffer(fs_info, search);
2201         if (eb) {
2202                 free_extent_buffer(eb);
2203                 return;
2204         }
2205
2206         target = search;
2207
2208         nritems = btrfs_header_nritems(node);
2209         nr = slot;
2210
2211         while (1) {
2212                 if (path->reada == READA_BACK) {
2213                         if (nr == 0)
2214                                 break;
2215                         nr--;
2216                 } else if (path->reada == READA_FORWARD) {
2217                         nr++;
2218                         if (nr >= nritems)
2219                                 break;
2220                 }
2221                 if (path->reada == READA_BACK && objectid) {
2222                         btrfs_node_key(node, &disk_key, nr);
2223                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2224                                 break;
2225                 }
2226                 search = btrfs_node_blockptr(node, nr);
2227                 if ((search <= target && target - search <= 65536) ||
2228                     (search > target && search - target <= 65536)) {
2229                         readahead_tree_block(fs_info, search);
2230                         nread += blocksize;
2231                 }
2232                 nscan++;
2233                 if ((nread > 65536 || nscan > 32))
2234                         break;
2235         }
2236 }
2237
2238 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2239                                        struct btrfs_path *path, int level)
2240 {
2241         int slot;
2242         int nritems;
2243         struct extent_buffer *parent;
2244         struct extent_buffer *eb;
2245         u64 gen;
2246         u64 block1 = 0;
2247         u64 block2 = 0;
2248
2249         parent = path->nodes[level + 1];
2250         if (!parent)
2251                 return;
2252
2253         nritems = btrfs_header_nritems(parent);
2254         slot = path->slots[level + 1];
2255
2256         if (slot > 0) {
2257                 block1 = btrfs_node_blockptr(parent, slot - 1);
2258                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2259                 eb = find_extent_buffer(fs_info, block1);
2260                 /*
2261                  * if we get -eagain from btrfs_buffer_uptodate, we
2262                  * don't want to return eagain here.  That will loop
2263                  * forever
2264                  */
2265                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2266                         block1 = 0;
2267                 free_extent_buffer(eb);
2268         }
2269         if (slot + 1 < nritems) {
2270                 block2 = btrfs_node_blockptr(parent, slot + 1);
2271                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2272                 eb = find_extent_buffer(fs_info, block2);
2273                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2274                         block2 = 0;
2275                 free_extent_buffer(eb);
2276         }
2277
2278         if (block1)
2279                 readahead_tree_block(fs_info, block1);
2280         if (block2)
2281                 readahead_tree_block(fs_info, block2);
2282 }
2283
2284
2285 /*
2286  * when we walk down the tree, it is usually safe to unlock the higher layers
2287  * in the tree.  The exceptions are when our path goes through slot 0, because
2288  * operations on the tree might require changing key pointers higher up in the
2289  * tree.
2290  *
2291  * callers might also have set path->keep_locks, which tells this code to keep
2292  * the lock if the path points to the last slot in the block.  This is part of
2293  * walking through the tree, and selecting the next slot in the higher block.
2294  *
2295  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2296  * if lowest_unlock is 1, level 0 won't be unlocked
2297  */
2298 static noinline void unlock_up(struct btrfs_path *path, int level,
2299                                int lowest_unlock, int min_write_lock_level,
2300                                int *write_lock_level)
2301 {
2302         int i;
2303         int skip_level = level;
2304         int no_skips = 0;
2305         struct extent_buffer *t;
2306
2307         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2308                 if (!path->nodes[i])
2309                         break;
2310                 if (!path->locks[i])
2311                         break;
2312                 if (!no_skips && path->slots[i] == 0) {
2313                         skip_level = i + 1;
2314                         continue;
2315                 }
2316                 if (!no_skips && path->keep_locks) {
2317                         u32 nritems;
2318                         t = path->nodes[i];
2319                         nritems = btrfs_header_nritems(t);
2320                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2321                                 skip_level = i + 1;
2322                                 continue;
2323                         }
2324                 }
2325                 if (skip_level < i && i >= lowest_unlock)
2326                         no_skips = 1;
2327
2328                 t = path->nodes[i];
2329                 if (i >= lowest_unlock && i > skip_level) {
2330                         btrfs_tree_unlock_rw(t, path->locks[i]);
2331                         path->locks[i] = 0;
2332                         if (write_lock_level &&
2333                             i > min_write_lock_level &&
2334                             i <= *write_lock_level) {
2335                                 *write_lock_level = i - 1;
2336                         }
2337                 }
2338         }
2339 }
2340
2341 /*
2342  * This releases any locks held in the path starting at level and
2343  * going all the way up to the root.
2344  *
2345  * btrfs_search_slot will keep the lock held on higher nodes in a few
2346  * corner cases, such as COW of the block at slot zero in the node.  This
2347  * ignores those rules, and it should only be called when there are no
2348  * more updates to be done higher up in the tree.
2349  */
2350 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2351 {
2352         int i;
2353
2354         if (path->keep_locks)
2355                 return;
2356
2357         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2358                 if (!path->nodes[i])
2359                         continue;
2360                 if (!path->locks[i])
2361                         continue;
2362                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2363                 path->locks[i] = 0;
2364         }
2365 }
2366
2367 /*
2368  * helper function for btrfs_search_slot.  The goal is to find a block
2369  * in cache without setting the path to blocking.  If we find the block
2370  * we return zero and the path is unchanged.
2371  *
2372  * If we can't find the block, we set the path blocking and do some
2373  * reada.  -EAGAIN is returned and the search must be repeated.
2374  */
2375 static int
2376 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2377                       struct extent_buffer **eb_ret, int level, int slot,
2378                       const struct btrfs_key *key)
2379 {
2380         struct btrfs_fs_info *fs_info = root->fs_info;
2381         u64 blocknr;
2382         u64 gen;
2383         struct extent_buffer *b = *eb_ret;
2384         struct extent_buffer *tmp;
2385         struct btrfs_key first_key;
2386         int ret;
2387         int parent_level;
2388
2389         blocknr = btrfs_node_blockptr(b, slot);
2390         gen = btrfs_node_ptr_generation(b, slot);
2391         parent_level = btrfs_header_level(b);
2392         btrfs_node_key_to_cpu(b, &first_key, slot);
2393
2394         tmp = find_extent_buffer(fs_info, blocknr);
2395         if (tmp) {
2396                 /* first we do an atomic uptodate check */
2397                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2398                         *eb_ret = tmp;
2399                         return 0;
2400                 }
2401
2402                 /* the pages were up to date, but we failed
2403                  * the generation number check.  Do a full
2404                  * read for the generation number that is correct.
2405                  * We must do this without dropping locks so
2406                  * we can trust our generation number
2407                  */
2408                 btrfs_set_path_blocking(p);
2409
2410                 /* now we're allowed to do a blocking uptodate check */
2411                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2412                 if (!ret) {
2413                         *eb_ret = tmp;
2414                         return 0;
2415                 }
2416                 free_extent_buffer(tmp);
2417                 btrfs_release_path(p);
2418                 return -EIO;
2419         }
2420
2421         /*
2422          * reduce lock contention at high levels
2423          * of the btree by dropping locks before
2424          * we read.  Don't release the lock on the current
2425          * level because we need to walk this node to figure
2426          * out which blocks to read.
2427          */
2428         btrfs_unlock_up_safe(p, level + 1);
2429         btrfs_set_path_blocking(p);
2430
2431         if (p->reada != READA_NONE)
2432                 reada_for_search(fs_info, p, level, slot, key->objectid);
2433
2434         ret = -EAGAIN;
2435         tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2436                               &first_key);
2437         if (!IS_ERR(tmp)) {
2438                 /*
2439                  * If the read above didn't mark this buffer up to date,
2440                  * it will never end up being up to date.  Set ret to EIO now
2441                  * and give up so that our caller doesn't loop forever
2442                  * on our EAGAINs.
2443                  */
2444                 if (!extent_buffer_uptodate(tmp))
2445                         ret = -EIO;
2446                 free_extent_buffer(tmp);
2447         } else {
2448                 ret = PTR_ERR(tmp);
2449         }
2450
2451         btrfs_release_path(p);
2452         return ret;
2453 }
2454
2455 /*
2456  * helper function for btrfs_search_slot.  This does all of the checks
2457  * for node-level blocks and does any balancing required based on
2458  * the ins_len.
2459  *
2460  * If no extra work was required, zero is returned.  If we had to
2461  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2462  * start over
2463  */
2464 static int
2465 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2466                        struct btrfs_root *root, struct btrfs_path *p,
2467                        struct extent_buffer *b, int level, int ins_len,
2468                        int *write_lock_level)
2469 {
2470         struct btrfs_fs_info *fs_info = root->fs_info;
2471         int ret;
2472
2473         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2474             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2475                 int sret;
2476
2477                 if (*write_lock_level < level + 1) {
2478                         *write_lock_level = level + 1;
2479                         btrfs_release_path(p);
2480                         goto again;
2481                 }
2482
2483                 btrfs_set_path_blocking(p);
2484                 reada_for_balance(fs_info, p, level);
2485                 sret = split_node(trans, root, p, level);
2486                 btrfs_clear_path_blocking(p, NULL, 0);
2487
2488                 BUG_ON(sret > 0);
2489                 if (sret) {
2490                         ret = sret;
2491                         goto done;
2492                 }
2493                 b = p->nodes[level];
2494         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2495                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2496                 int sret;
2497
2498                 if (*write_lock_level < level + 1) {
2499                         *write_lock_level = level + 1;
2500                         btrfs_release_path(p);
2501                         goto again;
2502                 }
2503
2504                 btrfs_set_path_blocking(p);
2505                 reada_for_balance(fs_info, p, level);
2506                 sret = balance_level(trans, root, p, level);
2507                 btrfs_clear_path_blocking(p, NULL, 0);
2508
2509                 if (sret) {
2510                         ret = sret;
2511                         goto done;
2512                 }
2513                 b = p->nodes[level];
2514                 if (!b) {
2515                         btrfs_release_path(p);
2516                         goto again;
2517                 }
2518                 BUG_ON(btrfs_header_nritems(b) == 1);
2519         }
2520         return 0;
2521
2522 again:
2523         ret = -EAGAIN;
2524 done:
2525         return ret;
2526 }
2527
2528 static void key_search_validate(struct extent_buffer *b,
2529                                 const struct btrfs_key *key,
2530                                 int level)
2531 {
2532 #ifdef CONFIG_BTRFS_ASSERT
2533         struct btrfs_disk_key disk_key;
2534
2535         btrfs_cpu_key_to_disk(&disk_key, key);
2536
2537         if (level == 0)
2538                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2539                     offsetof(struct btrfs_leaf, items[0].key),
2540                     sizeof(disk_key)));
2541         else
2542                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2543                     offsetof(struct btrfs_node, ptrs[0].key),
2544                     sizeof(disk_key)));
2545 #endif
2546 }
2547
2548 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2549                       int level, int *prev_cmp, int *slot)
2550 {
2551         if (*prev_cmp != 0) {
2552                 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2553                 return *prev_cmp;
2554         }
2555
2556         key_search_validate(b, key, level);
2557         *slot = 0;
2558
2559         return 0;
2560 }
2561
2562 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2563                 u64 iobjectid, u64 ioff, u8 key_type,
2564                 struct btrfs_key *found_key)
2565 {
2566         int ret;
2567         struct btrfs_key key;
2568         struct extent_buffer *eb;
2569
2570         ASSERT(path);
2571         ASSERT(found_key);
2572
2573         key.type = key_type;
2574         key.objectid = iobjectid;
2575         key.offset = ioff;
2576
2577         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2578         if (ret < 0)
2579                 return ret;
2580
2581         eb = path->nodes[0];
2582         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2583                 ret = btrfs_next_leaf(fs_root, path);
2584                 if (ret)
2585                         return ret;
2586                 eb = path->nodes[0];
2587         }
2588
2589         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2590         if (found_key->type != key.type ||
2591                         found_key->objectid != key.objectid)
2592                 return 1;
2593
2594         return 0;
2595 }
2596
2597 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2598                                                         struct btrfs_path *p,
2599                                                         int write_lock_level)
2600 {
2601         struct btrfs_fs_info *fs_info = root->fs_info;
2602         struct extent_buffer *b;
2603         int root_lock;
2604         int level = 0;
2605
2606         /* We try very hard to do read locks on the root */
2607         root_lock = BTRFS_READ_LOCK;
2608
2609         if (p->search_commit_root) {
2610                 /* The commit roots are read only so we always do read locks */
2611                 if (p->need_commit_sem)
2612                         down_read(&fs_info->commit_root_sem);
2613                 b = root->commit_root;
2614                 extent_buffer_get(b);
2615                 level = btrfs_header_level(b);
2616                 if (p->need_commit_sem)
2617                         up_read(&fs_info->commit_root_sem);
2618                 /*
2619                  * Ensure that all callers have set skip_locking when
2620                  * p->search_commit_root = 1.
2621                  */
2622                 ASSERT(p->skip_locking == 1);
2623
2624                 goto out;
2625         }
2626
2627         if (p->skip_locking) {
2628                 b = btrfs_root_node(root);
2629                 level = btrfs_header_level(b);
2630                 goto out;
2631         }
2632
2633         /*
2634          * If the level is set to maximum, we can skip trying to get the read
2635          * lock.
2636          */
2637         if (write_lock_level < BTRFS_MAX_LEVEL) {
2638                 /*
2639                  * We don't know the level of the root node until we actually
2640                  * have it read locked
2641                  */
2642                 b = btrfs_read_lock_root_node(root);
2643                 level = btrfs_header_level(b);
2644                 if (level > write_lock_level)
2645                         goto out;
2646
2647                 /* Whoops, must trade for write lock */
2648                 btrfs_tree_read_unlock(b);
2649                 free_extent_buffer(b);
2650         }
2651
2652         b = btrfs_lock_root_node(root);
2653         root_lock = BTRFS_WRITE_LOCK;
2654
2655         /* The level might have changed, check again */
2656         level = btrfs_header_level(b);
2657
2658 out:
2659         p->nodes[level] = b;
2660         if (!p->skip_locking)
2661                 p->locks[level] = root_lock;
2662         /*
2663          * Callers are responsible for dropping b's references.
2664          */
2665         return b;
2666 }
2667
2668
2669 /*
2670  * btrfs_search_slot - look for a key in a tree and perform necessary
2671  * modifications to preserve tree invariants.
2672  *
2673  * @trans:      Handle of transaction, used when modifying the tree
2674  * @p:          Holds all btree nodes along the search path
2675  * @root:       The root node of the tree
2676  * @key:        The key we are looking for
2677  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2678  *              deletions it's -1. 0 for plain searches
2679  * @cow:        boolean should CoW operations be performed. Must always be 1
2680  *              when modifying the tree.
2681  *
2682  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2683  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2684  *
2685  * If @key is found, 0 is returned and you can find the item in the leaf level
2686  * of the path (level 0)
2687  *
2688  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2689  * points to the slot where it should be inserted
2690  *
2691  * If an error is encountered while searching the tree a negative error number
2692  * is returned
2693  */
2694 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2695                       const struct btrfs_key *key, struct btrfs_path *p,
2696                       int ins_len, int cow)
2697 {
2698         struct btrfs_fs_info *fs_info = root->fs_info;
2699         struct extent_buffer *b;
2700         int slot;
2701         int ret;
2702         int err;
2703         int level;
2704         int lowest_unlock = 1;
2705         /* everything at write_lock_level or lower must be write locked */
2706         int write_lock_level = 0;
2707         u8 lowest_level = 0;
2708         int min_write_lock_level;
2709         int prev_cmp;
2710
2711         lowest_level = p->lowest_level;
2712         WARN_ON(lowest_level && ins_len > 0);
2713         WARN_ON(p->nodes[0] != NULL);
2714         BUG_ON(!cow && ins_len);
2715
2716         if (ins_len < 0) {
2717                 lowest_unlock = 2;
2718
2719                 /* when we are removing items, we might have to go up to level
2720                  * two as we update tree pointers  Make sure we keep write
2721                  * for those levels as well
2722                  */
2723                 write_lock_level = 2;
2724         } else if (ins_len > 0) {
2725                 /*
2726                  * for inserting items, make sure we have a write lock on
2727                  * level 1 so we can update keys
2728                  */
2729                 write_lock_level = 1;
2730         }
2731
2732         if (!cow)
2733                 write_lock_level = -1;
2734
2735         if (cow && (p->keep_locks || p->lowest_level))
2736                 write_lock_level = BTRFS_MAX_LEVEL;
2737
2738         min_write_lock_level = write_lock_level;
2739
2740 again:
2741         prev_cmp = -1;
2742         b = btrfs_search_slot_get_root(root, p, write_lock_level);
2743
2744         while (b) {
2745                 level = btrfs_header_level(b);
2746
2747                 /*
2748                  * setup the path here so we can release it under lock
2749                  * contention with the cow code
2750                  */
2751                 if (cow) {
2752                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2753
2754                         /*
2755                          * if we don't really need to cow this block
2756                          * then we don't want to set the path blocking,
2757                          * so we test it here
2758                          */
2759                         if (!should_cow_block(trans, root, b)) {
2760                                 trans->dirty = true;
2761                                 goto cow_done;
2762                         }
2763
2764                         /*
2765                          * must have write locks on this node and the
2766                          * parent
2767                          */
2768                         if (level > write_lock_level ||
2769                             (level + 1 > write_lock_level &&
2770                             level + 1 < BTRFS_MAX_LEVEL &&
2771                             p->nodes[level + 1])) {
2772                                 write_lock_level = level + 1;
2773                                 btrfs_release_path(p);
2774                                 goto again;
2775                         }
2776
2777                         btrfs_set_path_blocking(p);
2778                         if (last_level)
2779                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2780                                                       &b);
2781                         else
2782                                 err = btrfs_cow_block(trans, root, b,
2783                                                       p->nodes[level + 1],
2784                                                       p->slots[level + 1], &b);
2785                         if (err) {
2786                                 ret = err;
2787                                 goto done;
2788                         }
2789                 }
2790 cow_done:
2791                 p->nodes[level] = b;
2792                 btrfs_clear_path_blocking(p, NULL, 0);
2793
2794                 /*
2795                  * we have a lock on b and as long as we aren't changing
2796                  * the tree, there is no way to for the items in b to change.
2797                  * It is safe to drop the lock on our parent before we
2798                  * go through the expensive btree search on b.
2799                  *
2800                  * If we're inserting or deleting (ins_len != 0), then we might
2801                  * be changing slot zero, which may require changing the parent.
2802                  * So, we can't drop the lock until after we know which slot
2803                  * we're operating on.
2804                  */
2805                 if (!ins_len && !p->keep_locks) {
2806                         int u = level + 1;
2807
2808                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2809                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2810                                 p->locks[u] = 0;
2811                         }
2812                 }
2813
2814                 ret = key_search(b, key, level, &prev_cmp, &slot);
2815                 if (ret < 0)
2816                         goto done;
2817
2818                 if (level != 0) {
2819                         int dec = 0;
2820                         if (ret && slot > 0) {
2821                                 dec = 1;
2822                                 slot -= 1;
2823                         }
2824                         p->slots[level] = slot;
2825                         err = setup_nodes_for_search(trans, root, p, b, level,
2826                                              ins_len, &write_lock_level);
2827                         if (err == -EAGAIN)
2828                                 goto again;
2829                         if (err) {
2830                                 ret = err;
2831                                 goto done;
2832                         }
2833                         b = p->nodes[level];
2834                         slot = p->slots[level];
2835
2836                         /*
2837                          * slot 0 is special, if we change the key
2838                          * we have to update the parent pointer
2839                          * which means we must have a write lock
2840                          * on the parent
2841                          */
2842                         if (slot == 0 && ins_len &&
2843                             write_lock_level < level + 1) {
2844                                 write_lock_level = level + 1;
2845                                 btrfs_release_path(p);
2846                                 goto again;
2847                         }
2848
2849                         unlock_up(p, level, lowest_unlock,
2850                                   min_write_lock_level, &write_lock_level);
2851
2852                         if (level == lowest_level) {
2853                                 if (dec)
2854                                         p->slots[level]++;
2855                                 goto done;
2856                         }
2857
2858                         err = read_block_for_search(root, p, &b, level,
2859                                                     slot, key);
2860                         if (err == -EAGAIN)
2861                                 goto again;
2862                         if (err) {
2863                                 ret = err;
2864                                 goto done;
2865                         }
2866
2867                         if (!p->skip_locking) {
2868                                 level = btrfs_header_level(b);
2869                                 if (level <= write_lock_level) {
2870                                         err = btrfs_try_tree_write_lock(b);
2871                                         if (!err) {
2872                                                 btrfs_set_path_blocking(p);
2873                                                 btrfs_tree_lock(b);
2874                                                 btrfs_clear_path_blocking(p, b,
2875                                                                   BTRFS_WRITE_LOCK);
2876                                         }
2877                                         p->locks[level] = BTRFS_WRITE_LOCK;
2878                                 } else {
2879                                         err = btrfs_tree_read_lock_atomic(b);
2880                                         if (!err) {
2881                                                 btrfs_set_path_blocking(p);
2882                                                 btrfs_tree_read_lock(b);
2883                                                 btrfs_clear_path_blocking(p, b,
2884                                                                   BTRFS_READ_LOCK);
2885                                         }
2886                                         p->locks[level] = BTRFS_READ_LOCK;
2887                                 }
2888                                 p->nodes[level] = b;
2889                         }
2890                 } else {
2891                         p->slots[level] = slot;
2892                         if (ins_len > 0 &&
2893                             btrfs_leaf_free_space(fs_info, b) < ins_len) {
2894                                 if (write_lock_level < 1) {
2895                                         write_lock_level = 1;
2896                                         btrfs_release_path(p);
2897                                         goto again;
2898                                 }
2899
2900                                 btrfs_set_path_blocking(p);
2901                                 err = split_leaf(trans, root, key,
2902                                                  p, ins_len, ret == 0);
2903                                 btrfs_clear_path_blocking(p, NULL, 0);
2904
2905                                 BUG_ON(err > 0);
2906                                 if (err) {
2907                                         ret = err;
2908                                         goto done;
2909                                 }
2910                         }
2911                         if (!p->search_for_split)
2912                                 unlock_up(p, level, lowest_unlock,
2913                                           min_write_lock_level, &write_lock_level);
2914                         goto done;
2915                 }
2916         }
2917         ret = 1;
2918 done:
2919         /*
2920          * we don't really know what they plan on doing with the path
2921          * from here on, so for now just mark it as blocking
2922          */
2923         if (!p->leave_spinning)
2924                 btrfs_set_path_blocking(p);
2925         if (ret < 0 && !p->skip_release_on_error)
2926                 btrfs_release_path(p);
2927         return ret;
2928 }
2929
2930 /*
2931  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2932  * current state of the tree together with the operations recorded in the tree
2933  * modification log to search for the key in a previous version of this tree, as
2934  * denoted by the time_seq parameter.
2935  *
2936  * Naturally, there is no support for insert, delete or cow operations.
2937  *
2938  * The resulting path and return value will be set up as if we called
2939  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2940  */
2941 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2942                           struct btrfs_path *p, u64 time_seq)
2943 {
2944         struct btrfs_fs_info *fs_info = root->fs_info;
2945         struct extent_buffer *b;
2946         int slot;
2947         int ret;
2948         int err;
2949         int level;
2950         int lowest_unlock = 1;
2951         u8 lowest_level = 0;
2952         int prev_cmp = -1;
2953
2954         lowest_level = p->lowest_level;
2955         WARN_ON(p->nodes[0] != NULL);
2956
2957         if (p->search_commit_root) {
2958                 BUG_ON(time_seq);
2959                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2960         }
2961
2962 again:
2963         b = get_old_root(root, time_seq);
2964         level = btrfs_header_level(b);
2965         p->locks[level] = BTRFS_READ_LOCK;
2966
2967         while (b) {
2968                 level = btrfs_header_level(b);
2969                 p->nodes[level] = b;
2970                 btrfs_clear_path_blocking(p, NULL, 0);
2971
2972                 /*
2973                  * we have a lock on b and as long as we aren't changing
2974                  * the tree, there is no way to for the items in b to change.
2975                  * It is safe to drop the lock on our parent before we
2976                  * go through the expensive btree search on b.
2977                  */
2978                 btrfs_unlock_up_safe(p, level + 1);
2979
2980                 /*
2981                  * Since we can unwind ebs we want to do a real search every
2982                  * time.
2983                  */
2984                 prev_cmp = -1;
2985                 ret = key_search(b, key, level, &prev_cmp, &slot);
2986
2987                 if (level != 0) {
2988                         int dec = 0;
2989                         if (ret && slot > 0) {
2990                                 dec = 1;
2991                                 slot -= 1;
2992                         }
2993                         p->slots[level] = slot;
2994                         unlock_up(p, level, lowest_unlock, 0, NULL);
2995
2996                         if (level == lowest_level) {
2997                                 if (dec)
2998                                         p->slots[level]++;
2999                                 goto done;
3000                         }
3001
3002                         err = read_block_for_search(root, p, &b, level,
3003                                                     slot, key);
3004                         if (err == -EAGAIN)
3005                                 goto again;
3006                         if (err) {
3007                                 ret = err;
3008                                 goto done;
3009                         }
3010
3011                         level = btrfs_header_level(b);
3012                         err = btrfs_tree_read_lock_atomic(b);
3013                         if (!err) {
3014                                 btrfs_set_path_blocking(p);
3015                                 btrfs_tree_read_lock(b);
3016                                 btrfs_clear_path_blocking(p, b,
3017                                                           BTRFS_READ_LOCK);
3018                         }
3019                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3020                         if (!b) {
3021                                 ret = -ENOMEM;
3022                                 goto done;
3023                         }
3024                         p->locks[level] = BTRFS_READ_LOCK;
3025                         p->nodes[level] = b;
3026                 } else {
3027                         p->slots[level] = slot;
3028                         unlock_up(p, level, lowest_unlock, 0, NULL);
3029                         goto done;
3030                 }
3031         }
3032         ret = 1;
3033 done:
3034         if (!p->leave_spinning)
3035                 btrfs_set_path_blocking(p);
3036         if (ret < 0)
3037                 btrfs_release_path(p);
3038
3039         return ret;
3040 }
3041
3042 /*
3043  * helper to use instead of search slot if no exact match is needed but
3044  * instead the next or previous item should be returned.
3045  * When find_higher is true, the next higher item is returned, the next lower
3046  * otherwise.
3047  * When return_any and find_higher are both true, and no higher item is found,
3048  * return the next lower instead.
3049  * When return_any is true and find_higher is false, and no lower item is found,
3050  * return the next higher instead.
3051  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3052  * < 0 on error
3053  */
3054 int btrfs_search_slot_for_read(struct btrfs_root *root,
3055                                const struct btrfs_key *key,
3056                                struct btrfs_path *p, int find_higher,
3057                                int return_any)
3058 {
3059         int ret;
3060         struct extent_buffer *leaf;
3061
3062 again:
3063         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3064         if (ret <= 0)
3065                 return ret;
3066         /*
3067          * a return value of 1 means the path is at the position where the
3068          * item should be inserted. Normally this is the next bigger item,
3069          * but in case the previous item is the last in a leaf, path points
3070          * to the first free slot in the previous leaf, i.e. at an invalid
3071          * item.
3072          */
3073         leaf = p->nodes[0];
3074
3075         if (find_higher) {
3076                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3077                         ret = btrfs_next_leaf(root, p);
3078                         if (ret <= 0)
3079                                 return ret;
3080                         if (!return_any)
3081                                 return 1;
3082                         /*
3083                          * no higher item found, return the next
3084                          * lower instead
3085                          */
3086                         return_any = 0;
3087                         find_higher = 0;
3088                         btrfs_release_path(p);
3089                         goto again;
3090                 }
3091         } else {
3092                 if (p->slots[0] == 0) {
3093                         ret = btrfs_prev_leaf(root, p);
3094                         if (ret < 0)
3095                                 return ret;
3096                         if (!ret) {
3097                                 leaf = p->nodes[0];
3098                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3099                                         p->slots[0]--;
3100                                 return 0;
3101                         }
3102                         if (!return_any)
3103                                 return 1;
3104                         /*
3105                          * no lower item found, return the next
3106                          * higher instead
3107                          */
3108                         return_any = 0;
3109                         find_higher = 1;
3110                         btrfs_release_path(p);
3111                         goto again;
3112                 } else {
3113                         --p->slots[0];
3114                 }
3115         }
3116         return 0;
3117 }
3118
3119 /*
3120  * adjust the pointers going up the tree, starting at level
3121  * making sure the right key of each node is points to 'key'.
3122  * This is used after shifting pointers to the left, so it stops
3123  * fixing up pointers when a given leaf/node is not in slot 0 of the
3124  * higher levels
3125  *
3126  */
3127 static void fixup_low_keys(struct btrfs_path *path,
3128                            struct btrfs_disk_key *key, int level)
3129 {
3130         int i;
3131         struct extent_buffer *t;
3132         int ret;
3133
3134         for (i = level; i < BTRFS_MAX_LEVEL; i++) {